SLAM and Map Learning using Hybrid Semantic Graph Optimization *

Abstract

Visual Simultaneous Localization and Mapping using budget-grade cameras only faces the challenges of continuous drifts that accumulate with time. While loop closure techniques mitigate the effects, they are applicable only when the robot completes a loop, which is a rarity in everyday navigation. The motion blur and smaller resolution of budget cameras further reduce the accuracy of SLAM. In this paper, we aim to solve the problem of active drift correction for a low-cost robot to solve autonomous navigation using the semantic map. Semantic maps have been used previously for re-localization but are useful only when the semantic maps themselves are highly accurate which is not realizable for budget robots. The semantic maps also face problems of correspondence matching in areas rich with recurrent semantics. To alleviate the same effects, the robot performs SLAM using a hybrid graph optimization consisting of semantic points whose pose is obtained from the semantic map database, and the non-semantic point features. The semantic map corrects for the drift, while the non-semantic features apply local smoothing that helps in mitigating the errors of the semantic map. They also apply robustness against errors in correspondence matching. The semantic graph may itself have errors, which are hence learned with time as the robot navigates. The robot adds new semantic objects into the database if it observes them, while the robot also mends the position based on the new observations. The initial semantic map is made using images captured by a camera on a few known poses, based on which it adds the observed semantics.